ShipDesk.devblog
the blog

Prompting AI coding assistants: give context, not wishes

George · 2 min read

The short answer: tell your assistant what you want to achieve and why, point it at the files and patterns that matter, say what it must not change, and ask for small steps you can check. A vague wish like "make the dashboard better" gets a guess. A clear task with context gets code you can ship.

AI coding assistants are now part of most solo developers' day. The difference between frustrating and fantastic results is often not the model, but how the work is described.

Describe the goal, not just the change

"Add a button" is a change. "Users keep asking how to export their data. Add an export button on the invoices page that downloads a CSV of the invoices they can see" is a goal. With the goal, the assistant can make better small decisions on its own, such as which data to include and where the button belongs.

Give it the right context

  • Point to examples in your code. "Follow the pattern used in the projects page" keeps new code consistent with old code.
  • Mention the constraints: the framework version, the database, the libraries you already use.
  • Share the error in full when fixing a bug: the message, the stack trace, and what you did when it happened.
  • Keep a project notes file that your assistant reads, with how to run tests, conventions and things to avoid. Many tools support this.

Set limits

Say what must stay the same: "Do not change the public API", "Do not add new dependencies", "Only touch the files in this folder". Assistants are eager to help and sometimes improve things you did not ask about. Limits keep each change small and reviewable.

Work in small steps

Ask for one change at a time and check it before the next. For bigger features, ask the assistant to propose a plan first, agree on it, then build it step by step. If a step goes wrong, it is easy to undo one small change and hard to untangle a large one.

Ask it to check its own work

Ask the assistant to run the tests, the type checker and the linter, and to tell you what it could not verify. Then check the parts that matter yourself, especially anything involving money, permissions or deleting data.

When it keeps getting it wrong

If you have corrected the same mistake twice, stop and change the approach. Give a concrete example of the expected result, split the task smaller, or start a fresh session with a clearer description. Repeating "no, fix it" rarely works.

The short version

  • Explain the goal and why, not only the change.
  • Point to files, patterns and constraints.
  • Say what must not change.
  • Work in small, checkable steps.
  • Verify the important parts yourself.

If you build with Claude Code or Cursor, you can even launch from your editor: ShipDesk has an MCP server that turns your project into a launch draft.

launch from your AI assistant